The invention discloses a vehicle detection method based on semantic reasoning and time-space memory updating. The method comprises the steps of firstly, establishing a 3D traffic scene model, and associating an input image with the 3D space to determine a 3D candidate vehicle search area; binarizing the vehicle search area, and generating candidate vehicles by adopting a fast scan line algorithm; performing semantic reasoning and cross validation between a 2D image and the 3D space, and filtering the candidate vehicles by sufficiently using prior knowledge, a constraint relation between vehicles, dependency between things and vehicle appearance information; and improving the vehicle detection accuracy and robustness by using time-space consistency. The candidate vehicle generating process is performed in the 3D space, and this operation can amplify distant vehicles and effectively improve the detection rate of distant small-sized vehicles; the time-space memory updating mechanism can reduce missing detection of vehicles having high confidence, which is helpful to improve the detection rate; and the semantic reasoning mechanism reduces the false detection rate of the method.